Knowledge Cutoffs and Live Search: How to Check AI Answers

A knowledge cutoff limits a model's training knowledge; live search can supply newer material at answer time. Search availability does not establish that a particular response searched, fetched your page or cited it. Check the response evidence and the exact product settings.

A simple answer pipeline

A useful model is: question → available model knowledge and tools → optional retrieval → generated answer → optional citations. This is a conceptual guide, not a claim that every vendor exposes the same internal architecture.

Training knowledge resides in model parameters. Retrieval adds material to the current context from a search index, a document store or a fetched URL. Grounding means tying an answer to supplied evidence; it is not a guarantee that every sentence is correct. Citation is the visible attribution. None of these steps proves the next one happened.

A dated article can therefore appear in a search-grounded answer even if it was published after the underlying model's cutoff. Conversely, a recently trained model can give a stale answer. The date of training is not a live-web freshness guarantee.

Product, API and settings matter

The following is a documentation review checked on 21 October 2026, not a hands-on quality comparison or a current model leaderboard.

A consumer chat interface is a separate surface from an API. Its mode, account state, personalisation and available tools may differ. A tracker using an API does not automatically reproduce what a reader sees in a consumer app. Compare providers only after documenting those differences; our tools comparison helps identify the measurement method.

For a specific cutoff, open the vendor's model documentation for the exact model identifier used in the capture. Do not infer a shared cutoff for an entire company or product family. If the model is not disclosed, record unknown.

How to check an answer's sources

Here is a worked inspection procedure, not a reported live query result:

1. Save the question and time in UTC, plus the interface, mode, model if disclosed, location and search settings.
2. Inspect the response's search/tool trace or source panel. Search enabled without a search call is not evidence of retrieval.
3. Open one cited URL. Find the passage that supports the sentence, check its publication/review date and note redirects or unavailable pages.
4. Record separate fields for search observed, URL cited, claim supported and page fetched if the tool explicitly reports that action. A citation alone cannot prove a fresh fetch of the origin.
5. Mark an error, blocked source or missing trace as unknown/unavailable, not as a negative result. A valid answer without your citation is a measured zero only within that capture's defined denominator.

For example, if a response cites an old cached guide, the presence of a citation does not make its price current. Checking the source can reveal that mismatch without making a second model call. We have not executed a new cross-provider demonstration for this update; that would require a dated protocol and, where relevant, an approved API budget.

What this means for a publisher

Publish useful text at stable, reachable URLs and make each material claim inspectable. Dates tell the reader when a claim was reviewed; they do not force an engine to retrieve it. A crawler request establishes access, not inclusion in training or a generated answer. A brand mention without a link is different from a citation.

Keep your reporting layers separate: captured citations, referral visits, search impressions and verified crawler requests. Our Lab displays the boundaries and case study explains why historical results are not interchangeable with later coverage.

Review method and limitations

We reviewed the primary documentation linked above and corrected unsupported generalisations. AI assisted with this documentation review and rewrite. We did not measure comparative model behaviour, prove source preference or repeat the old experiment. Source documentation describes supported mechanisms; actual responses still need inspection.

Read What is GEO?, the methods taxonomy, citation rate and Methodology for definitions and a proportionate measurement plan. More explainers are in Foundations.